External Damage Risk Detection of Transmission Lines Using E-OHEM Enhanced Faster R-CNN

Abstract

External damage risk detection of transmission lines suffers from complex background information and small-sized objects, thereby leading to numerous false negatives and false positives. Although faster R-CNN with OHEM algorithm can achieve impressive results for the task of external damage risk detection, data imbalance remains an obstacle when applying OHEM. In object detection, one of the major challenges is how to learn an effective model from imbalanced data. To address this issue, we propose a variant of OHEM named Enhance Online Hard Example Mining (E-OHEM) algorithm for external damage risk detection of transmission lines, which can not only effectively learn more hard examples but also avoid data imbalance. Moreover, we fine-tune the learned model by selecting negative examples around false positives at the RPN stage. The fine-tuning model outperforms the original Faster R-CNN model by 0.5% on the mAP value on the PASCAL VOC2007 dataset. The application of E-OHEM algorithm and fine-tuned model render our detector more robust and efficient for the task of external damage risk detection. Meanwhile, our model makes a significant improvement on recognition accuracy.